Numpyro determistic not tracking variables during MCMC

numpyro
1

Hi everyone,

I have some confusion over how numpyro determistic works. I have a fairly complex model, but, the key part is this:

        with numpyro.plate("strain", self.n_variants):
            strain_auc = numpyro.sample("strain_auc", dist.LogNormal(0, 0.5))

        strain_auc_to_save = numpyro.deterministic("strain_auc_to_save", strain_auc)

when I run MCMC, and plot the posterior distribution of the two variables, what I get is this:

image
image1852×1628 213 KB

Any idea what might be going on? Is determistic not doing what I expect it to do?

2

I should add: if I do map estimation (using numpyro_ext) - then the deterministic node and the samples node are exactly the same, so this seems to be something that’s triggered by doing MCMC sampling with NUTS.

3

What a bug - could you try

numpyro.deterministic("strain_auc_to_save", strain_auc + 1.0)

?